Stony Brook University researchers Billy Tsouvalas and Nick Nikiforakis found that bots pull secrets from public container images soon after posting. They presented the work at the ACM CCS conference, per BleepingComputer on 17 November. Over four months, they put 300 decoy images on eight registries, holding fake AWS, OpenAI and Claude keys.
BleepingComputer reported that one planted AWS key was first used 7.86 seconds after its image went live on Docker Hub. That was the fastest case; the median first call came at about 51 seconds, and all 100 AWS keys were used within 15 minutes. The AWS keys drew over 140,000 calls from more than 5,000 IP addresses, nearly all on Docker Hub. The OpenAI and Claude test accounts were drained in short bursts. Scanners rarely ran the images. They mostly unpacked the layers, the file bundles inside an image. Tsouvalas said Docker only acknowledged the May report, and an August test showed no change.
Teams that publish public images should treat the push as the moment a secret is exposed, since reading files needs no running container. These keys were decoys, but the report says revoking real leaked keys takes significant work and coordination. The researchers urge Docker Hub to scan for secrets before images go live and to limit its feed of new images.